slide-builder（幻灯片生成器）

Slide Builder

Transform essay-to-speech output into complete, presentation-ready slide decks with multiple output format support (HTML, Remotion video, PowerPoint).

When to Use

Invoke when user:

• Has essay-to-speech output and wants slides
• Says "create slides from this talk track"
• Needs to "build a presentation" from spoken content
• Wants to convert a talk track to video format
• Uses /slide-builder command
• Asks for "presentation slides" from transformed essay content

Prerequisites

Input required: Output from the essay-to-speech skill containing:

• ### Original sections (verbatim essay text)
• ### Talk Track sections with semantic tags
• ### Images sections with ratings (USE/ADAPT/RECREATE/SKIP)
• ### Slide Ideas suggestions

Core Process

1. Parse Essay-to-Speech Output

Extract structured data from each section:

Section → {
  title: string,
  original: string,
  talkTrack: TaggedContent[],
  images: ImageAssessment[],
  slideIdeas: string[]
}

Semantic tags to identify:

• [HOOK] - Opening attention-grabber → Title/hook slide
• [KEY_POINT] - Core argument → Statement slide
• [EVIDENCE] - Data/proof → Data visualization slide
• [STORY] - Narrative → Story/quote slide
• [TRANSITION] - Bridge → Section divider or no slide
• [CALLBACK] - Reference → Recap element
• [LANDING] - Conclusion → Summary slide
• [CTA] - Call to action → Action slide

2. Plan Slide Deck

Map semantic tags to slides:

Slide count heuristic:

• 1-2 slides per [KEY_POINT]
• 1 slide per [EVIDENCE] block
• Section dividers are optional (skip for tight decks)
• Target: 1 slide per 45-60 seconds of speaking

3. Handle Images

Process image assessments from essay-to-speech:

For RECREATE images: Generate a Nano Banana prompt following these guidelines:

• 16:9 aspect ratio for slides
• Clear, simple compositions
• Large readable text/labels
• Brand colors if specified

Example RECREATE prompt:

"Clean horizontal bar chart comparing 5 items, minimal style,
white background, teal (#557373) bars, large bold labels,
no gridlines, presentation-ready, 16:9 aspect ratio"

4. Generate Output

Output Format: Talk Track v5

The primary output format for presentations.

Structure

# Talk Track v5 frontmatter
version: 5
title: "Presentation Title"
subtitle: "Optional Subtitle"
author: "Presenter Name"
date: "2025-01-15"
target_minutes: 15
audio_voice: "af_heart"
brand:
  primary: "#557373"
  background: "#F2EFEA"
  text: "#0D0D0D"
sections:
  - id: opening
    name: "Opening"
    color: "#557373"
  - id: problem
    name: "The Problem"
    color: "#6B8E6B"
  - id: solution
    name: "The Solution"
    color: "#C4785A"
  - id: closing
    name: "Closing"
    color: "#557373"
---

## Slides

| # | Slug | Title | Image | Section |
|---|------|-------|-------|---------|
| 1 | hook | The Question | hook.png | opening |
| 2 | problem-1 | What's Broken | problem-chart.png | problem |
| 3 | evidence | The Data | evidence.png | problem |
| 4 | solution | A New Approach | solution.png | solution |
| 5 | action | Your Next Step | cta.png | closing |

---

## [hook] The Question

![The Question](images/hook.png)

<!-- AUDIO -->
[HOOK] Let me ask you something that might change how you think about this entire problem...

What if everything you believed was based on outdated assumptions?
<!-- /AUDIO -->

**Speaker Notes:**
- Pause after the question
- Make eye contact with audience
- Let the tension build

---

## [problem-1] What's Broken

![What's Broken](images/problem-chart.png)

<!-- AUDIO -->
[KEY_POINT] The current approach fails in three critical ways.

[EVIDENCE] First, efficiency drops by 40% when teams scale past 10 people. Second, communication overhead grows exponentially. Third, institutional knowledge gets siloed.
<!-- /AUDIO -->

**Speaker Notes:**
- Point to chart as you mention each stat
- Emphasize "exponentially"

---

Format Rules

1. YAML Frontmatter - Metadata, timing, voice, sections
2. Slide Index Table - Quick reference for all slides
3. Individual Slides - Each with:

- H2 header: ## [slug] Title - Image reference (if applicable) - <!-- AUDIO --> block with talk track - **Speaker Notes:** for presenter context

Audio Block Format

Content between <!-- AUDIO --> and <!-- /AUDIO --> is:

• Read aloud by TTS engines
• Preserves semantic tags for timing hints
• Excludes speaker notes and visual descriptions

Timing Calculation

Estimate duration based on word count:

• Speaking rate: 130-150 words/minute
• Add 2-3 seconds per slide transition
• Add pause time for [PAUSE] markers

Voice Options

Development: Kokoro TTS (via claude-speak)

• Local, free, fast iteration
• Invoke: /claude-speak skill with audio block text
• Voice: af_heart (default) or specify in frontmatter
• Output: Local audio file per slide or full presentation

Production: ElevenLabs v3

• Word-level timestamps for precise video sync
• Higher quality for final distribution
• Requires ElevenLabs API key
• Output: Audio + JSON timing data

See references/voice-options.md for full configuration.

Alternative Output Formats

HTML Slide Engine

Static HTML presentation with:

• Keyboard navigation (arrows, space)
• Speaker notes toggle (N key)
• Timer display
• Print to PDF support

See references/html-engine.md for template.

Remotion Video

Export to React-based video for:

• YouTube/social publishing
• Embedded animations
• Precise audio sync with timestamps

See references/remotion-video.md for project setup.

Workflow

Standard Flow

essay-to-speech output
        ↓
   [slide-builder]
        ↓
   Talk Track v5 (.md)
        ↓
   ┌─────┼─────┐
   ↓     ↓     ↓
  HTML  Video  Audio

Quick Start

1. Input: Provide essay-to-speech output
2. Review plan: Claude proposes slide structure
3. Confirm or adjust: Modify slide count, sections, visuals
4. Generate: Claude outputs Talk Track v5 markdown
5. Images: Generate RECREATE images via Nano Banana
6. Audio: Generate voice via claude-speak or ElevenLabs
7. Render: Export to HTML, video, or both

Best Practices

Slide Design Principles

1. One idea per slide - Split dense content
2. 6 words or less on screen - The rest is spoken
3. High contrast - Readable from back row
4. Consistent visual language - Same fonts, colors, style
5. Images > bullet points - Visual storytelling wins

Talk Track Integration

1. Audio is king - Slides support speech, not replace it
2. Match pacing - Visual changes align with spoken transitions
3. Build reveals - Don't show everything at once
4. Breathing room - Not every sentence needs a slide change

Image Guidance

For RECREATE images, always specify:

• Aspect ratio (16:9 for slides)
• Style (clean, minimal, professional)
• Key data to visualize
• What to AVOID (clutter, small text, decorative elements)

What This Skill Does NOT Do

• Edit or create original essay content (that's essay-to-speech)
• Design custom graphics (use Nano Banana for that)
• Record actual audio (use claude-speak or ElevenLabs)
• Render final video (use Remotion or video editor)
• Create PowerPoint/Keynote files directly (exports markdown)

Integration

Upstream:

• essay-to-speech - Provides structured input

Downstream:

• nano-banana - Generates RECREATE images
• claude-speak - Generates audio narration
• veo3-prompter - Creates video segments (if needed)

References

• references/talk-track-v5.md - Complete format specification
• references/html-engine.md - Static HTML slide player
• references/remotion-video.md - React video export setup
• references/voice-options.md - TTS configuration and comparison
• references/image-handling.md - Full image processing workflow
• references/examples.md - Complete input→output examples